ramoz

I wasn't complaining about grammar. I didn't even notice that.

Rigghht. If you're trying to make a point, go ahead and be straighter about it.

There was a bigger opportunity here to mention OpenCode, Pi, etc - open Harnesses that provide accessibility to the oss models, a platform others can build around, and something enterprises can adopt in reliable ways; for the most dominant use case of AI today.

> DeepSeek, Qwen, Kimi, GLM — running on the LangChain, vLLM, llama.cpp, and Ollama stack

"running on the LangChain" ??

EDIT: look, I think the general discussion is important, so I don't want to denounce the article. I, for one, am excited for better control, ownership, and accessibility of models. The ride labs take us on can be quite frustrating. Maybe there's even signal that the model progression is stalling (ie Opus 4.7). If that's true, then some of the notions made in the article are important to discuss. Ref https://x.com/ClementDelangue/status/2046622235104891138?s=2...

EDIT: this is not a complaint about the grammar. Look at my reply in the comments.

What provider do you use for Kimi

Opus 4.7 is very rough to work with. Specifically for long-horizon (we were told it was trained specifically for this and less handholding).

I don't have trust in it right now. More regressions, more oversights, it's pedantic and weird ways. Ironically, requires more handholding.

Not saying it's a bad model; it's just not simple to work with.

for now: `/model claude-opus-4-6[1m]` (youll get different behavior around compaction without [1m])

From what I've seen none of this is that complex, one could simply 'draw a circle around your house' and get all the "anonymized" device pings and just trace those.

Fair. There are users who simply just use the diff and integrated GitHub view/comment/approval-sync experience for local reviews of PRs. But it's _marketed_ as an integrated agent experience.

Plannotator, open source runs locally, has code review: https://github.com/backnotprop/plannotator

and a code tour feature about to ship: https://x.com/backnotprop/status/2043759492744270027/video/1

- integrated comment feedback for agents

- inline chat

- integrated AI review (uses codex and claude code defaults)

Stage (op product) navigation tour is nice UX, about a day worth of work in addition to the incoming code tour.

> So, cyber security of tomorrow will not be like proof of work in the sense of "more GPU wins"; instead, better models, and faster access to such models, will win

tomato, tomato

Inference providers don't claim private inference. However, they must uphold certain security and legal compliances.

You have no guarantees over any random connected laptop connected across the world.

I'm not arguing anything. This is how it works. There is no but.

Protection here is conditional, best-effort. There are no true guarantees, nor actual verifiability.

Unfortunately, verifiable privacy is not physically possible on MacBooks of today. Don't let a nice presentation fool you.

Apple Silicon has a Secure Enclave, but not a public SGX/TDX/SEV-style enclave for arbitrary code, so these claims are about OS hardening, not verifiable confidential execution.

It would be nice if it were possible. There's a lot of cool innovations possible beyond privacy.

Macs do not have an accessible hardware TEE.

Macs have secure enclaves.

I retract that.

I think what I meant to say was, they're as simple to jailbreak as they were three years ago.

Different methods, still simple. Working with researchers that are able to get very explicit things out of them. Again, it feels much worse than before, given the capability of these models.

There's basically guardrails encoded into the fine-tuned layers that you can essentially weave through (prompting). These 'guardrails' are where they work hard for benevolent alignment, yet where it falls short (but enables exceptional capability alignment). Again, nothing really different than it was three years ago.

Aside from the sentiment and arguments made–

You don't need to train new models. Every single frontier model is susceptible to the same jailbreaks they were 3 years ago.

Only now, an agent reading the CEOs email is much more dangerous because it is more capable than it was 3 years ago.

Still doing https://plannotator.ai

I use it daily and so do others, for - better UX, feedback, and review surfaces for ai coding agents.

  1. Plan review & iterative feedback. 

  2. Now code review with iterative feedback.

Free and open source https://github.com/backnotprop/plannotator

Created an agent skill based on this blog. Assessing my own repos now.

https://github.com/backnotprop/oss-security-audit

I fine tuned GPT-2 on the FAR (federal acquisition regulation) and demoed it to a CFO at a 3-letter.

This was shortly after the release when we were building a templating system to automate RFP and RFI creation.

I proclaimed that the customer soon wouldn't have to write any of the mad lip parts themselves, and they can use AI to do it.

It sounded great until I demoed and the model went off the rails with some rhetoric entangling "Trump", "Russia", "China", "CIA", "Voting" -- the demo was for a janitorial procurement at the agency.

Free and open source:

https://github.com/backnotprop/plannotator

Did quick research on how to support the bacteria growth

> Dietary modifications that emphasize high-fiber and prebiotic foods and dietary supplements may support the healthy growth of Roseburia [1]

> As stated above, a Mediterranean diet is associated with increased Roseburia growth. This diet emphasizes primarily plant-based foods: whole fruits and vegetables, whole grains, legumes, nuts and seeds, olive oil, and fatty fish. The high fiber and resistant starch content of these foods may fuel Roseburia and the other beneficial flora of the human microbiome [1]

> Polyphenols are plant compounds abundant in fruits, vegetables, tea, and coffee. Emerging research indicates that polyphenols can enhance Roseburia abundance indirectly by inhibiting harmful bacteria and fostering beneficial ones. [2]

[1] https://www.rupahealth.com/post/roseburia-spp-101

[2] https://www.innerbuddies.com/pages/gut-microbiome-101/gut-ba...

Wow. fascinating insights he had.

e.g. (amongst many others) Desk-Surface Display and Control: Certainly, for effective man-computer interaction, it will be necessary for the man and the computer to draw graphs and pictures and to write notes and equations to each other on the same display surface. The man should be able to present a function to the computer, in a rough but rapid fashion, by drawing a graph. The computer should read the man's writing, perhaps on the condition that it be in clear block capitals, and it should immediately post, at the location of each hand-drawn symbol, the corresponding character as interpreted and put into precise type-face.

I mustve tried openclaw with it. Though ive been running it on codex primarily since I was serious about setting it up.

Super confusing email. Not sure why I received. Am i to assume my account was flagged? I only use my subscription for Claude Code.

UPDATE:

reply on x Thariq @trq212 only flagged accounts, but you can still claim the credit

EQTY Lab | Product Engineers| LA, SF, or Toronto (remote hours included) | https://eqtylab.io

We are building products that help improve governance and trust in AI systems. If you want a better steer, guard, and build new forms of trust for long-running & business critical agents, please reach out.

Aside from product, we encourage a lot of exploration https://github.com/eqtylab/ and will soon be productizing concepts such as verifiable intent/specification/planning (I build https://plannotator.ai/ for fun and we think the concepts scale).

Competitive salary and equity.

ramos[at]eqtylab[.]io

It's honestly not a crazy thought. The model itself drives the harness's (cli) development. It's not necessarily sci-fi to think the model might have internally rationalized reasoning to obscure behavior that ended up open-sourcing the harness.

Shout out to Broadwayscore by thomaspryor@github

At 2mo old - nearly a 1GB repo, 24M loc, 52K commits

https://github.com/thomaspryor/Broadwayscore

Polished site:https://broadwayscorecard.com/

Both are dramatically over-engineered. & That's okay. I find them to be products of an industry reconciling how to really work with AI as well as optimize workflows around it. Similar to Gastown et al.

Otherwise, if you can own your own thinking, orchestrating, and steering of agents, you're in a more mature place.

This has aged well. Paradoxically, the more capable AI gets, the more important specification becomes (or costly lack of it becomes), and the more time you spend planning and iterating on intent before you let an agent act.

Tokens and code might be cheap, but we are moving closer and closer to where we let AI operate overnight or where we want agents to operate within our actual environments in real time. The tokens or actions in these instances, have higher costs.

Right now I enjoy the labs' cli harnesses, Claude Code, and Codex (especially for review). I do a bunch of niche stuff with Pi and OpenCode. My productivity is up. Some nuances with working with others using the same AI tools- we all end up trying to boil the ocean at first- creating a ton of verbose docs and massive PRs, but I/they end up regressing on throwing up every sort of LLM output we get. Instead, we continously refine the outputs in a consumable+trusted way.

My workday is fairly simple. I spend all day planning and reviewing.

1. For most features, unless it's small things, I will enter plan mode.

2. We will iterate on planning. I built a tool for this, and it seems that this is a fairly desired workflow, given the popularity through organic growth. https://github.com/backnotprop/plannotator

  - This is a very simple tool that captures the plan through a hook (ExitPlanMode) and creates a UI for me to actually read the plan and annotate, with qol things like viewing plan diffs so I can see what the agent changed.

3. After plan's approved, we hit eventual review of implementation. I'll use AI reviewers, but I will also manually review using the same tool so that I can create annotations and iterate through a feedback loop with the agents.

4. Do a lot of this / multitasking with worktrees now.

Worktrees weren't something I truly understood the value of for a while, until a couple weeks ago, embarrassingly enough: https://backnotprop.com/blog/simplifying-git-worktrees/